FinInteract (Frontier Pilot): leaderboard
The FinInteract protocol (Wang et al., 2026) run as a preliminary pilot on a stratified 50-instance subset: the agent may search the filings and ask a simulated user yes-or-no questions, and must answer under the intended interpretation. Accuracy in percent on the pilot subset, which is not comparable with the full 173-instance board; higher is better.
Metric: Accuracy (%; the stratified 50-instance pilot subset run for the cross-vendor panel; ReAct agent in the standard full-interaction mode (A+S+I): it may search the filing corpus, ask the simulated user (GPT-5, answering yes, no or I don't know from the hidden disambiguating context) yes-or-no questions, and answer; a final answer counts as correct only if it matches the intended interpretation, graded by GPT-4o-mini with a 1% numeric tolerance). Source: arxiv.org. Saturation forecast: Around 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 5 | 44 | #108 |
| 2 | GLM-5.2 | 40 | #137 |
| 3 | Kimi K3 | 34 | #53 |
| 4 | Grok 4.5 | 32 | #78 |
| 5 | GPT-5.6 Sol | 30 | #29 |
| 6 | Gemini 3.5 Flash | 30 | #98 |
| 7 | DeepSeek V4 Flash | 24 | #254 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=fininteract-frontier-pilot · How It Works · Data refreshed daily, snapshot 2026-11-12.